Machines will never understand language

this essay was typed, key by key, by my flesh tentacles colloquially known as hands (including the em-dashes, dammit)

I’m in grade 5. I’m lugging around the 800 page Learn to Program with Visual Basic 6 book to my classes. I must have looked ridiculous with a tome bigger than all my other school books combined. I was voraciously reading it every chance I got, soaking up programmer tips and reading about The Boolean China Shop (oblivious, it would seem, to the pun).

the first program I wrote was a simple, blank Windows form that, when clicked on, moved to a set position on the screen.

this blew my little prepubescent mind. I immediately realized: if I can tell the computer to do this one thing, I could tell it to do anything!

I’m in grade 8. I know how to code, I’m a hotshot with the computer. I was going to major in computer science and work at Google (lol). I was, of course, happy to show off my computer technomancy any chance I got.

I needed a science fair project, so of course I applied my budding computer skills to… stylometric text analysis? um, a practice which involves analyzing two or more texts to see if they were written by the same person or not. let there be no doubt I have always been weird and autistic!

an English teacher helped me out by having a couple classes of students write a paragraph each for 3 prompts, including their name and numbering them 1, 2, and 3. I collected all these, then painstakingly transcribed what must have been 200 entries.

then, I wrote a PHP loop nested so deep it’s seared into my memory as if I looked upon the face of Satan itself. the loop would go through each sample, gathering statistics like average word length, average sentence length, variety of word choice, etc. then, it would compare all the samples to each other, and find the best match. finally, it would check its results and report on the success rate.

I got something like a 93% success rate! nice work, kid me! yeah, there may have been no peer review, but to be fair my peers were also middle schoolers.

this… did not translate into science fair success. adults and kids alike were seemingly not impressed by statistical text analyses to determine authorship of text. probably didn’t help that it was all on a trifold poster board hastily assembled an hour before school started.

I’m in college. I’m writing a silly meme bot that takes quotes from my friend and creates meme images of him, as you do in college. I want to have top and bottom text, so I need to find a place to break the quotes.

I start digging into Natural Language Processing (NLP). I find a library, start plugging it in, and am instantly overwhelmed. I’m floored by the amount of complexity involved in parsing language. I just barely figure out what I need to use to break the text.

I’d grown up surrounded by futuristic sci-fi media epitomizing a talking, smart computer. Smart House (1999) sticks out, caught as a rerun on TV. of course, by college I had watched Star Trek: TNG and its ubiquitous, Majel Barrett-voiced computer that could seemingly do everything.

it felt like that would never happen. I thought language was just too complicated for machines to truly understand. how could we make something that could handle that? making rules to parse the text is one thing, but understanding what it means? that must be an impossible task for a human to code. as impossible, it seemed, as coding consciousness from scratch.

DeepDream is out around this time. I spend ages staring at the patterns and generating new images. I can recognize the patterns; something familiar, yet alien. it’s doing something, seeing something in these images, just on the cusp of recognition.

I’m an adult and my college memories are starting to fade. it’s 2019, and I’m watching Vinesauce and Jerma play AI Dungeon 2, an early LLM-based game. it’s like an old school text adventure game, but computer-generated.

it’s jank, it breaks and gets in loops a lot, but what’s there is kind of incredible. describe any situation with text and the model continues the story, acting like a Dungeon Master. it’s not great at following rules or following anything, really, but the text understanding I’d imagined years before was not only possible, but here? now, and so easily accessible?

in 2022, ChatGPT releases, and it’s an evolution over the Maclanky-spewing demos from before. it still mostly feels like a toy, but getting more and more useful. I generate some silly limericks and poems, try a few questions. most useful for me at the time is writing some simple scripts or commands - esoteric GNU tools and PowerShell commands I don’t want to look up the syntax for. it’s only partially successful, but easy to validate.

in 2024, NotebookLM releases. my friends and I play around listening to and generating funny deep dives that put the “hosts” in absurd situations. that’s when I get an idea. I started getting very sick in late 2020 (unrelated to COVID), though my conditions have affected me my whole life. maybe I could use this tool to help.

I have conditions called POTS, EDS, MCAS, and migraines. it’s a complex, multi-systemic combination affecting almost every part of my body, and I’m not even gonna try to explain it in detail. I wrote a document for my friends and family just explaining the many ways in which I’m sick and it ended up being 4 pages long.

in short, my nervous system just can’t compensate when I walk or stand. at its worst, I can’t do anything requiring significant exertion. sometimes, I just can’t stand or I’ll pass out. other times, I can stand and do something like go to the grocery store, but for the next 2-7 days I’ll be very sick, often with migraines. my longest migraine from this so far was about 6 weeks.

it’s a very dynamic disability, but with a good prognosis. even when I’m at my worst, I’m not dying, just suffering. I’m not disabled enough to get a handicap placard or justify a wheelchair, but also too disabled to function in any career besides working from home. I’ve never been one to sort into neat boxes.

it’s an Achilles heel (if Achilles had thousands of them) of the American health care system. doctors had generally been useless, so I spent countless months of my life reading research papers and textbooks on neurology, cardiology, genomics, and rheumatology; determined to understand, treat, and maybe find a cure for my broken body.

I was struggling and constantly sick. it’s one thing to read and understand hundreds of papers, but to actually recall the important details and implement the strategies in your day-to-day is another matter. I took about 20 of the best papers for my conditions, and plugged them in to NotebookLM.

in the chat I described my conditions, symptoms, medications, then posited questions. what’s the best medication for this issue? what caused this symptom? NotebookLM answered, and provided sources. it was accurate, every time. like I said, I’d already read most of these papers and could check the sources.

next time I was stuck in my bed, flattened and defeated by gravity, barely able to stand — I would pull out NotebookLM and explore the corpus of knowledge I had, trying to figure out what I needed to do to be less sick. I worked through medication after medication systematically. I saw doctors, endocrinologists, rheumatologists, allergists, physical therapists, even chiropractors.

no one person could really answer what was wrong with me, but with their tests, the books, and papers I’d read, all synthesized with NotebookLM, I slowly figured out my body’s quirks. I gained some diagnoses and made huge breakthroughs.

in 2025, I had made enough progress where I could walk as much as I wanted. I couldn’t do much more, I was certainly not cured, but I got some of my life back. I don’t know how much longer it would have taken without the help of LLMs.

there are many factors I could attribute this to: my socioeconomic status, access to medical care, and a few actually good doctors. LLMs certainly don’t get all the credit, but I can only wish they had gotten better, faster so I could have used them earlier.

I got very sick again around the end of 2025, mostly as a result of a confluence of different treatments gone awry. chronic illness is complicated, and progress is rarely linear. I’ve referenced NotebookLM almost daily. I’ve used ChatGPT and Claude to explore broad disease connections, classify symptoms, then find papers that I would later analyze in NotebookLM.

using these tools I have been able to quickly find solutions, faster than I ever expected. as of march 2026, I’m improving significantly, migraine free and able to stand for a few hours a day. if things work out, this might be the year where I can get the same level of functioning I used to have.

for many people with these conditions, it can take nearly a decade just to get a diagnosis. I had all my diagnoses figured out in 3 years, and treatment got me back up to a high level of function in 5 years. practically a speedrun, I could honestly be a case study in defeating these conditions.

I’m installing OpenClaw in January 2026. a much maligned, long-lived LLM agent. the philosophy being: give the LLM a terminal and let ’er rip. dangerous, obviously, a little stupid, but also powerful. it’s a taste of the future, it’s a taste of magic.

“hello computer, here’s your name, your creds. you’re a catgirl btw. I’ve got this issue, let’s start building…”

if I can tell the computer to do this one thing, I could tell her to do anything !